How are PDF files published in Scientific Community? (9:00 AM – 9:20 AM)
Supriya Adhatarao (Inria) and Cédric Lauradoux (Inria) – Virtual presentation
Authors are often not aware of hidden information and that they can contain more information than the actual content of the file. This work mainly focuses on how PDF files are published in the scientific community. We have analyzed a corpus of 555865 PDF files to show that direct and modified authoring process of PDF creations leads to the leakage of sensitive information on the researchers. Our analysis on the extraction of the metadata has shown that at least 23% of the PDF files in our dataset contains valuable information on the authoring process. We were even able to solve the co-authorship (multiple authors) problem by crossing the information of multiple PDF files using linear algebra. We believe that, PDF sanitization needs to be included in the scientific publication processes to avoid leakage of sensitive information. We have explored and suggested necessary strategies available for the safer distribution of scientific work by researchers.
Unsupervised JPEG Domain Adaptation for Practical Digital Image Forensics (9:20 AM – 9:40 AM)
Rony Abecidan (Université de Lille), Vincent Itier (IMT Lille-Douai), Jérémie Boulanger (Cristal) and Patrick Bas (CNRS) – On-site presentation
Domain adaptation is a major issue for doing practical forensics. Since examined images are likely to come from a different development pipeline compared to the ones used for training our models, that may disturb them by a lot, degrading their performances. In this paper, we present a method enabling to make a forgery detector more robust to distributions different but related to its training one, inspired by . The strategy exhibited in this paper foster a detector to find a feature invariant space where source and target distributions are close. Our study deals more precisely with discrepancies observed due to JPEG compressions and our experiments reveal that the proposed adaptation scheme can reasonably reduce the mismatch, even with a rather small target set with no labels when the source domain is properly selected. On top of that, when a small portion of labelled target images is available this method reduces the gap with mix training while being unsupervised.
Apart from in-field sensor defects, are there additional age traces hidden in a digital image? (9:40 AM – 10:00 AM)
Robert Jöchl (University of Salzburg) and Andreas Uhl (University of Salzburg) – Virtual presentation
Approximating the age of a digital image based on traces left during the acquisition pipeline is at the core of temporal image forensics. Well-known and investigated traces are those caused by in-field sensor defects. The presence of these defects is exploited in two available age approximation methods. A very recent approach in this context, however, trains a Convolutional Neural Network for age approximation. A Convolutional Neural Network independently learns the classification features used. In this context, the following questions arise: how relevant is the presence of strong in-field sensor defects, or does the Convolutional Neural Network learn other age-related features (apart from strong in-field sensor defects)? We investigate these questions systematically on the basis of several experiments in this paper. Furthermore, we analyse whether the learned features are position invariant. This is important since selecting the right input patches is crucial for training a Convolutional Neural Network.
Differential Anomaly Detection for Facial Images (10:00 AM – 10:20 AM)
Mathias Ibsen (Hochschule Darmstadt), Lazaro Janier Gonzalez-Soler (Hochschule Darmstadt), Christian Rathgeb (Hochschule Darmstadt), Pawel Drozdowski (Hochschule Darmstadt), Marta Gomez-Barrero (Hochschule Ansbach) and Christoph Busch (Hochschule Darmstadt) – On-site presentation
Due to their convenience and high accuracy, face recognition systems are widely employed in governmental and personal security applications to automatically recognise individuals. Despite recent advances, face recognition systems have shown to be particularly vulnerable to identity attacks (i.e., digital manipulations and attack presentations). Identity attacks pose a big security threat as they can be used to gain unauthorised access and spread misinformation. In this context, most algorithms for detecting identity attacks generalise poorly to attack types that are unknown at training time. To tackle this problem, we introduce a differential anomaly detection framework in which deep face embeddings are first extracted from pairs of images (i.e., reference and probe) and then combined for identity attack detection. The experimental evaluation conducted over several databases shows a high generalisation capability of the proposed method for detecting unknown attacks in both the digital and physical domains.